Discovering Scaling Exponents with Physics-Informed Müntz–Szász Networks — reproduction

OpenReview PVaFEuNnsD

ClaimVerdictHeadline
1 trainable exponents, two-timescalereproduced0.61% vs 25.0% with exponents frozen (41×)
2 270° wedge recovers μ = 2/3reproduced6.2e−5% error vs paper's 0.009%
3 40-config wedge benchmarkreproduced100% / 0.008% vs naive 70% / 1.10% (135×)
4 Theorem 4.5 error scalingfalsifiedΔ⁻² holds (−1.928); √N fails (slope 0.0000)
5 14.6% → 0.009% improvementfalsifiednaive is 0.60%, not 14.6%
6 Rayleigh limit Δ ≥ 0.1falsified10/10 at Δ=0.05; oracle resolves Δ=0.02